Dashboard

How to Mark Up AI API Costs You Resell

A flat multiple on token cost looks safe and quietly loses money, because the top one percent of users cost far more than the blended average suggests.

Manuele Estivo
Manuele Estivo
Growth & SEO Lead
8 October 20261 min read

If you build a product that calls a model on your customer's behalf, you are reselling AI API access whether you describe it that way or not. The question is how much to mark it up. A flat multiple on your token cost is the obvious answer and it is the one that quietly loses money, because usage in a customer base is not evenly distributed.

Here is the arithmetic that matters, and the formula that tells you when a plan is underwater.

Start with cost per active customer, not cost per token

Token prices are the input. They are not the number you price against. The number you price against is what a customer costs you in a month, and that depends almost entirely on how often they use the feature.

Take a product with a $29 monthly plan that includes an AI assistant. Suppose each assistant interaction costs you 4,000 input tokens and 600 output tokens. On a model priced at $0.10 per million input and $0.50 per million output, that is $0.0004 plus $0.0003, so roughly $0.0007 per interaction.

That looks free. At ten interactions a month it is free. The problem is the shape of the distribution.

Usage tier

Share of customers

Interactions per month

Monthly model cost

Margin on $29

Light

60%

20

$0.01

$28.99

Normal

30%

200

$0.14

$28.86

Heavy

9%

3,000

$2.10

$26.90

Extreme

1%

60,000

$42.00

minus $13.00

The blended cost per customer here is about $0.80, which is 2.8% of revenue and looks completely safe. But 1% of your customers are each losing you $13 a month, and that 1% is not random. It is your most engaged users, the ones who integrate you into a daily workflow, the ones who tell other people about you, and the ones most likely to be on an annual contract you cannot reprice until renewal.

The break-even formula

For any plan, the usage level at which a customer stops being profitable is:

break_even_calls = (plan_price - non_AI_cost_to_serve) / cost_per_call

In the example above, with $29 revenue, say $3 of non-AI cost to serve (hosting, support, payment fees), and $0.0007 per call:

(29 - 3) / 0.0007 = 37,142 calls per month

Now you have a number you can actually manage against. Anything under about 37,000 calls a month is profitable on this plan. Your extreme tier at 60,000 is not. The useful move is to find where your real customers sit against that line, not to guess a markup.

Run this once per plan. If your break-even point is a number no real customer could plausibly reach, your pricing is safe and you can stop worrying. If a handful of live accounts are already at half of it, you have a problem arriving in about six months.

Three ways to mark up AI API costs

Bundled with a generous cap. Include AI usage in the plan price, set a cap high enough that 99% of customers never see it, and make the cap visible in your pricing table. This is the right default for most products. It keeps the pricing page simple and it bounds your downside. The cap is not there to make money, it is there to stop one account from eating a plan's margin.

Metered passthrough with a multiple. You charge cost times a multiple, typically between 1.5x and 3x. This works when usage varies by an order of magnitude between customers and they already understand that. The multiple needs to cover more than tokens: failed calls you retry, the requests you make for your own quality checks, the support burden of explaining a bill, and the price risk of a provider changing rates. A 1.2x multiple is not a margin, it is a rounding error with extra steps.

Credits. The customer buys a pool and spends it. Credits decouple your pricing from provider pricing, which is their real advantage: when model prices drop, as they did again on 7 October 2026 when Anthropic cut Haiku prices by around 75%, your credit price does not have to move. The cost is that customers hate credits that expire and hate not knowing what a credit buys.

Build the price cut into the plan

Model prices have fallen repeatedly through 2026. That is good for your margin and it is a trap if you price as though it will continue.

Set your markup so the business works at today's prices, not at the price you expect in six months. If prices fall further, that is margin you did not plan for, which is the right way round. The alternative, pricing thin on the assumption of future cuts, means a provider holding prices flat for two quarters turns into a crisis.

The related decision is whether to pass cuts on, which is a positioning question rather than a maths one. We covered it in should you pass an AI price cut to your client.

What to do on Monday

  1. Calculate cost per call for your single highest-volume feature, using real token counts from logs rather than estimates.

  2. Compute break-even calls per plan with the formula above.

  3. Sort your accounts by usage and find where the top 1% sit against that number.

  4. If any live account is past break-even, decide now whether it gets a cap, a higher tier, or a conversation. Doing nothing is a decision to subsidise them indefinitely.

  5. Set a cap on every plan, even one nobody will hit. It is the cheapest insurance available.

For the mechanics of actually charging for usage, see how to add usage-based pricing to an AI app and usage-based versus flat-rate AI pricing. If you are billing a client rather than running a product, how to bill clients for AI API usage covers that case, and forecasting your AI bill is the input to all of it. The wider picture sits in our guide to AI monetization strategies.

Questions people ask

What is a normal markup on resold AI API costs?

Between 1.5x and 3x raw token cost when you meter it, though most products bundle instead and manage the risk with a cap. Below 1.5x rarely covers retries, overheads and price risk.

Should I show customers the underlying model cost?

Only if you are billing passthrough plus a fee, where transparency is the point. If you bundle, exposing token costs invites customers to price-shop your margin against the provider's public rates.

How do I stop one customer destroying my margin?

A hard cap per plan, with a clear upgrade path when they hit it. Caps are far easier to raise for a good customer than to introduce after the fact.

What happens to my pricing when model prices drop again?

Nothing, if you priced for today. A credit system or a bundled plan absorbs the cut as margin. Passthrough pricing hands it to the customer automatically.

Do I need to tell customers which model I use?

Not usually, and naming it commits you to it. A model change clause in your terms is the safer route.

How did this land?

About the author

Manuele Estivo
Manuele Estivo

Growth & SEO Lead

Manuele covers distribution: SEO, content strategy, and how AI-built products find their first thousand users. He tests everything he recommends.

Share

Get the next post in your inbox

One email a month. Product updates, engineering posts, and the best of Built with Swarmz.

I agree to receive emails about AI building tips and Swarmz product news. Unsubscribe any time.